“They can be matched or beat in performance by tweaked conventional CNNs: A ConvNet for the 2020s”
A ConvNet for the 2020s
把 ResNet 逐步现代化为 ConvNeXt,纯卷积网络追平视觉 Transformer
论文重新审视设计空间,检验纯卷积网络能达到什么水平:将标准 ResNet 逐步朝视觉 Transformer 的设计“现代化”,找出沿途导致性能差异的若干关键组件,得到名为 ConvNeXt 的纯卷积模型家族。ConvNeXt 完全由标准卷积模块构成,在准确率和可扩展性上与 Transformer 相比毫不逊色,ImageNet top-1 准确率达到 87.8%,在 COCO 检测和 ADE20K 分割上超过 Swin Transformer,同时保持了标准卷积网络的简洁和高效。
HN 上的人在争论视觉 Transformer 与 CNN 孰优孰劣时引用它,作为用现代训练技巧和深度卷积就能让 CNN 追平甚至超过 ViT 的证据。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“Using more modern training techniques along with depthwise convolutions results in equal if not better performance on vision tasks.”
“For example, in "A ConvNet for the 2020s"[1], the authors are able to scale up ConvNets to the sizes of ViTs.”
“Consider A ConvNet for the 2020s where the paper says”
“[2] A ConvNet for the 2020s and ConvNeXt V2”
“the ConvNext paper showed that applying those same tricks to convolutional networks can fully close the gap.”
- A ConvNet for the 2020s ↗carbocation · 2022-01-11 · 1 分
- A ConvNet for the 2020s ↗tboerstad · 2022-01-13 · 2 分
- A ConvNet for the 2020s ↗laybak · 2024-05-25 · 18 分